Capability matrix
Raw benchmark parity rarely decides the pick — tooling, open weights, and provider reliability usually do.
Claude Opus 5 leads on 12 of 12 shared benchmarks.
| Dimension | Claude Opus 5 | GLM 5.3 |
|---|---|---|
| GPQA Diamond | 93.4% | 88.1% |
| HLE | 56.3% | 42% |
| Terminal-Bench 2.1 | 89.1% | 88.2% |
| Terminal-Bench 3.0 | 43.3% | 28.3% |
| Terminal-Bench 4.0 | 51.8% | 37.9% |
| DeepSWE v1.1 | 74% | 66.9% |
| ProgramBench | 37% | 19% |
| NL2Repo-Bench | 75.3% | 58% |
| CyberGym | — | 84.5% |
| ExploitGym | 22.1% | 15% |
| HLE (w/tools) | 63.6% | 62.5% |
| Automation-Bench | 50.3% | 48.8% |
| Agents' Last Exam | 28.6% | 28.5% |
| Chartography (w/tools) | 84% | — |
| BabyVision (w/tools) | 94.1% | — |
| ZeroBench-main (w/tools) | 52% | — |
| Blended price | $15.00 /1M | — |
| Full priceInput · Output · Cached input | $5.00 · $25.00 · $0.500 | — |
| Context window | 1,000K | — |
Adjust the monthly volume — every model recomputes instantly.
Estimates use list prices and ignore cached input, batch discounts, and off-peak rates.